Why Science? | AI-for-Science Initiative

NRF’s 16 June 2026 announcement describes a national programme designed to transform the pace and scope of scientific discovery. The amount belongs to the initiative, not to any single student or project. Eight inaugural Challenge projects mark the launch; current awards and future calls should be checked through official channels.

AI-for-Science joins computational reasoning with domain science, so both the model and the physical or biological meaning of its output must be testable. The chapter develops programme literacy. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Record the amount, date, owner and scope in separate fields. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A programme budget is not an individual grant entitlement.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Next chapter

Machine-learning expertise and scientific expertise answer different parts of the problem. AI teams can design models and scalable workflows; domain scientists define meaningful variables, mechanisms, constraints and experiments. Collaboration should begin with a shared question and vocabulary rather than passing data across a disciplinary wall.

Faster screening can expand the number of hypotheses considered, but experimental validation, data quality and causal reasoning still determine whether a result deserves scientific confidence. The chapter develops interdisciplinary collaboration. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Create a bilingual glossary linking computational terms to domain meaning. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Cross-disciplinary labels do not guarantee integrated work.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The official release says Challenge grants support larger projects developing AI methods and tools for specific use areas, while Catalytic grants support smaller exploratory ideas at the frontier of science. Portfolio design can protect both coordinated ambition and creative testing. Readers should not infer unpublished eligibility or award terms.

The initiative’s idea of bilingual scientists is practical: researchers need enough AI fluency to formulate and audit models, and enough domain depth to recognise impossible or misleading outputs. The chapter develops portfolio reasoning. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Compare the evidence milestone suitable for a large challenge and a small exploratory study. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Only current official calls define application rules.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

NRF uses “bilingual” for researchers fluent in AI and a scientific field. This fluency includes translating domain needs into computational tasks, recognising invalid assumptions, communicating model limits and designing experiments that discriminate between explanations. It does not require one person to replace every specialist.

Responsible acceleration means measuring discovery quality, reproducibility, compute cost and scientific usefulness rather than celebrating speed alone. The chapter develops bilingual research literacy. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Write one domain question, model task and experimental check as a linked trio. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Fluency grows through teams, mentoring and practice.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Starting with a fashionable architecture can produce impressive outputs that do not resolve the scientific uncertainty. Define the mechanism, target quantity, prediction horizon, decision and acceptable error. Then compare simple baselines with more complex models. Complexity earns its place through better evidence.

AI-for-Science joins computational reasoning with domain science, so both the model and the physical or biological meaning of its output must be testable. The chapter develops problem specification. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Create a question-to-metric brief before choosing an algorithm. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A high benchmark score may not answer the scientific question.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Record where data came from, how samples were selected, which instruments or simulations produced them, what preprocessing changed and whose populations or conditions are missing. Large datasets can preserve systematic bias at scale. Metadata and version control let later researchers reproduce and challenge the work.

Faster screening can expand the number of hypotheses considered, but experimental validation, data quality and causal reasoning still determine whether a result deserves scientific confidence. The chapter develops data provenance. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Build a data card with origin, measurement, exclusions, permissions and limitations. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Access and reuse follow consent, contracts and applicable law.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Physical laws, conservation rules, chemical structure and biological knowledge can guide learning or rule out impossible outputs. The official portfolio includes physics-informed and knowledge-guided approaches. Constraints should be justified and tested because an incorrect prior can also steer a model away from reality.

The initiative’s idea of bilingual scientists is practical: researchers need enough AI fluency to formulate and audit models, and enough domain depth to recognise impossible or misleading outputs. The chapter develops constraint-aware modelling. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: List hard constraints, soft assumptions and ways each could be falsified. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Domain knowledge is a hypothesis input, not automatic truth.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Compare against existing methods, simple predictors and expert workflows using data not seen during training. Report uncertainty, subgroup performance and failure conditions. An improvement should matter in the units and decisions of the field, not only in an abstract score.

Responsible acceleration means measuring discovery quality, reproducibility, compute cost and scientific usefulness rather than celebrating speed alone. The chapter develops model evaluation. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Design an evaluation table with baseline, metric, split, uncertainty and practical threshold. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Test leakage or cherry-picked metrics can invalidate a result.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

AI may propose materials, molecules, mechanisms or operating conditions, but a prospective experiment tests whether the prediction survives contact with the world. Automated laboratories can speed this cycle. Keep failed experiments because they reveal model blind spots and improve the next iteration.

AI-for-Science joins computational reasoning with domain science, so both the model and the physical or biological meaning of its output must be testable. The chapter develops experimental validation. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Plan one prospective test whose outcome could change the model. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A generated candidate is not a discovered or safe solution.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Record code versions, dependencies, random seeds, hardware, prompts, model checkpoints and evaluation procedures where permitted. If data cannot be shared, consider documented access, synthetic examples or executable workflows. Reproduction should test scientific conclusions, not only rerun a script.

Faster screening can expand the number of hypotheses considered, but experimental validation, data quality and causal reasoning still determine whether a result deserves scientific confidence. The chapter develops computational reproducibility. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Create a reproducibility manifest and independent rerun plan. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Security, privacy and IP can restrict release.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The official annex lists projects in energy interfaces, genomics, program reasoning, materials synthesis, mRNA vaccines, agricultural digital twins, surface science and full blood-count analysis. Each has a different data-generating system and validation pathway. The list illustrates scope rather than proving future outcomes.

The initiative’s idea of bilingual scientists is practical: researchers need enough AI fluency to formulate and audit models, and enough domain depth to recognise impossible or misleading outputs. The chapter develops portfolio literacy. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Group the eight projects by question, data, model and real-world test. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Project aims remain aims until evidence is published.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Projects on interfaces, materials foundries and catalysts combine models with robotic or high-throughput experiments. Synthesis routes, defects and operating conditions determine whether a theoretical candidate can exist and perform. Record unsuccessful recipes because manufacturability is part of the discovery.

Responsible acceleration means measuring discovery quality, reproducibility, compute cost and scientific usefulness rather than celebrating speed alone. The chapter develops materials informatics. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Trace one candidate from representation to synthesis, characterisation and function. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Predicted properties do not establish practical materials.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Genomics, mRNA design and blood-count projects work with complex biological data. Define population, sample handling, outcome, confounding and clinical context; protect privacy and avoid treating associations as diagnoses. Any healthcare use requires suitable validation, oversight and regulation beyond research model performance.

AI-for-Science joins computational reasoning with domain science, so both the model and the physical or biological meaning of its output must be testable. The chapter develops biomedical AI literacy. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Create a population-and-use boundary statement for a biomedical model. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Research output is not medical advice or clinical approval.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The portfolio includes tools for reasoning about programs, specifications and proofs. Code that looks plausible can still be unsafe. Formal verification, tests and human review address different assurance layers; teams should define which property is proved and which remains empirical.

Faster screening can expand the number of hypotheses considered, but experimental validation, data quality and causal reasoning still determine whether a result deserves scientific confidence. The chapter develops software assurance. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Separate specification, proof obligation, tests and deployment monitoring. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A proof covers only the stated model and assumptions.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Larger models can consume substantial compute, time and energy. Report resources, compare efficiency and ask whether a smaller method answers the question. Shared computing infrastructure, scheduling and carbon-aware choices can widen access while supporting reproducibility.

The initiative’s idea of bilingual scientists is practical: researchers need enough AI fluency to formulate and audit models, and enough domain depth to recognise impossible or misleading outputs. The chapter develops compute stewardship. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Add compute, energy proxy and performance-per-resource measures to evaluation. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: More computation is not automatically better science.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Strong foundations in mathematics, statistics, programming and one scientific domain help, but communication, ethics, data care and experimental thinking are equally important. Build small projects that connect a model to a real measurement and document limitations. Use current official course and career sources rather than promising one route.

Responsible acceleration means measuring discovery quality, reproducibility, compute cost and scientific usefulness rather than celebrating speed alone. The chapter develops career preparation. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Create a six-week mini-project with data provenance, baseline and a physical or expert check. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A subject combination does not guarantee a research career.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Archive the official programme edition, project question, domain assumptions, data versions, models, compute, evaluations, experiments, failures, decisions, contributions and released outcomes. The record should let another team see what AI accelerated, what science constrained and what remains unknown.

AI-for-Science joins computational reasoning with domain science, so both the model and the physical or biological meaning of its output must be testable. The chapter develops evidence governance. Keep the source, decision context, evidence, feedback and revision together so another reader can see how the conclusion was reached. The NRF announcement defines the initiative and inaugural projects; current grant calls, host institutions, project teams, ethics bodies and data owners control participation, methods, access and results.

Put this into practice by turning the idea into a small, reviewable action: Build a model-experiment decision log with dated links and owners. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Retention and disclosure follow project, legal and institutional rules.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter